The number on the homepage is a probability, not a prediction. It is the model's estimate of the likelihood that the Islamic Republic undergoes a full regime change within the next 12 months, computed from seven underlying conditions — military, economic, social, political. A high reading means those conditions are stacked against the regime; it does not mean collapse is imminent on any particular date. Every headline figure on this site is computed end-to-end from the pillar scores using the formulas below — enforced in code at build time, not by procedure (see the Integrity Log). No hand-set numbers, no hidden adjustments: feed the model identical scores and you will get identical outputs.
As of Day 136 (July 13, 2026), the headline is the probability of full regime change within 12 months, given conditions as scored today. Attaching an explicit window makes the number scoreable: when a window closes, the forecast can be graded against reality, and we commit to publishing that grade regardless of the result. Two derived companions on the homepage — 90-day and 6-month — assume the risk is spread evenly across the year (constant hazard), a crude assumption we state rather than hide; collapse cascades are not evenly spread, and if the hazard were twice as concentrated in the next 90 days the 90-day companion would read roughly half again higher than shown. Read the number as a pressure gauge expressed on a probability scale: when conditions change, the number changes. Points on the chart before Day 136 were published without a window and should be read as unwindowed structural snapshots — they cannot be scored, which is part of why the window now exists.
How it would resolve. RC-1 resolves yes if the office of Supreme Leader is abolished or its holder loses effective command of the armed forces for 30+ consecutive days, or if the Islamic Republic's constitution is suspended or replaced. Any Transition additionally counts a managed succession outside the Assembly of Experts process, an open military takeover, or a constitutional rupture that removes the clerical veto structure. The window and these criteria together make the headline falsifiable — the grade will be published either way.
Resolution baselines (stipulated July 20, 2026). Three clarifications, published before any window closes so the criteria cannot be lawyered after the fact. First, each daily reading is a forward-looking forecast: it covers qualifying events initiating in the 365 days after that reading's date, measured against the state of affairs on that date. Second, for RC-1's command clause, the office-holder is deemed to hold effective command as of each reading's date — the clause resolves yes only on a subsequent, sustained 30-day loss; the incumbent's limited public visibility since March 2026 does not by itself qualify, and a scenario in which the office and its holder persist as figurehead under an open military takeover resolves Any Transition, not RC-1. Third, the March 2026 succession does not resolve Any Transition: it predates every windowed reading and was conducted through the Assembly of Experts process, however contested — “outside the Assembly process” means a future succession imposed without a formal Assembly vote. Any Transition carries the same 12-month window as the headline, measured from each reading's date; its first windowed reading is July 13, 2026.
Grading protocol. The scoring rule is the Brier score. Every daily RC-1 and Any Transition reading published from July 13, 2026 onward is graded at the close of its own 12-month window against the criteria and baselines above, and the running ledger is published as grades accrue — the first windows close in July 2027. The 90-day and 6-month companions are deterministic transforms of the headline, not independent forecasts, so they enter no separate ledger; at interim checkpoints (the first: about October 11, 2026) we will report whether a qualifying event occurred and how the derived companion fared, labeled as an indicator of the same underlying judgment. Resolution is adjudicated against reporting from at least two major international news agencies or an equivalent authoritative record, and the adjudication reasoning is published with the grade — regardless of the result.
The model decomposes regime stability into seven structural pillars, each scored 0–100 from open-source evidence, with weights that sum to 1.00. Four pillars measure collapse pressure directly: the more degraded Iran's military, the more fragile its economy, the more mobilized the population, and the more hostile the external environment, the higher the score and the more pressure on the regime. The other three are inverted — they measure defensive strength. A high score on security apparatus cohesion means the IRGC and Basij are still holding the line; a high score on information control means the regime is still setting the narrative; a high score on adaptive capacity means the leadership is still maneuvering effectively. For those three, the formula flips the score (100 − sᵢ) before aggregation. Security cohesion carries the heaviest weight at 25 percent, because in the historical cases the model was checked against (see Historical Grounding), the loyalty of the coercive apparatus was the last thing to break.
Editorial rule: each event is scored under exactly one primary pillar — a strike that kills a commander is military degradation; a failed succession rite is adaptive capacity; a desertion is security cohesion. Cross-references in prose are fine, but no event is double-counted. Each pillar's public scoring rubric (0–100 anchors) is published on the Pillars page.
The headline number is produced in two steps.
Step 1 — Compute the Collapse Pressure Index (CPI). The seven pillar scores are combined into a single index:
CPI = 100 · Π (sᵢ*/100)^wᵢ
For the four direct pillars, sᵢ* = sᵢ. For the three inverted pillars (P2 security cohesion, P6 information control, P7 adaptive capacity), sᵢ* = 100 − sᵢ, so regime strength reduces pressure rather than adding to it. The seven adjusted scores are combined as a weighted geometric mean (since Day 136; a weighted arithmetic sum before that). The geometric form is conjunctive: collapse pressure requires all conditions together, so a single strong regime defense drags the index down more than the old additive sum allowed, and pressure in one pillar can no longer fully substitute for another's — the lesson of Iraq 1991 (see Historical Grounding). The change introduced no new parameters: the weights and both logistic curves are unchanged. One structural property to note plainly: because the index is a product, a single pillar at its best-for-regime extreme (an effective score of zero) drives the whole index — and the headline — to zero. That conjunctive floor is intended (no amount of pressure elsewhere substitutes for a fully intact pillar), but it means that in that corner the model asserts survival with certainty, which is stronger than evidence could ever support. The result is a single number between 0 and 100. Today's value: 82.5.
Step 2 — Map CPI to probabilities via two independent logistic functions:
P(RC-1) = 100 / (1 + exp(−0.2 · (CPI − 83)))
P(Any Transition) = 100 / (1 + exp(−0.15 · (CPI − 70)))
RC-1 isolates full regime change. Any Transition is the broader category — managed succession, military takeover, constitutional rupture — and therefore sits above RC-1 at every CPI value. The midpoints (83, 70) and slopes (0.2, 0.15) were set by structured expert judgment, checked for consistency against historical cases of authoritarian failure (see Historical Grounding).
Two implications worth flagging. First, because the curve is steep near the threshold, small movements in pillar scores can shift the headline probability by several points. That is a feature: authoritarian systems fail in non-linear cascades, not smooth declines. The Sensitivity section below quantifies this number by number. Second, the model has no opinion about who or what replaces the current regime — it only measures whether the current arrangement can hold.
The headline is no more precise than its inputs. Each pillar is scored by human judgment from open-source evidence, and each score carries an observability uncertainty (one standard deviation, in score points) that is published alongside the model itself:
| Pillar | σ | Why |
|---|---|---|
| Military Degradation | 4 | Strike results and launch rates are partially verifiable from open sources |
| Security Apparatus Cohesion | 8 | Loyalty inside the coercive apparatus is the least observable quantity in the model |
| Economic Fragility | 5 | Prices and shipping are visible; the state's internal fiscal condition is not |
| Popular Mobilization | 6 | Blackouts and repression suppress protest visibility |
| External Environment | 3 | Diplomacy, strikes and shipping are publicly observable and multi-sourced |
| Information & Narrative Control | 5 | Broadcast and blackout status are observable; their reach and effect less so |
| Regime Adaptive Capacity | 6 | Elite decision-making is inferred from outcomes, not observed directly |
These propagate through the geometric index by the delta method — each pillar contributes tᵢ = CPI·wᵢ·σᵢ/sᵢ*, and σ_CPI = √Σtᵢ² (today 3.24) — and the same logistic curve that produces the headline is evaluated at CPI ± 1.28·σ. Assuming approximately normal scoring error, the monotone curve gives approximately 80% coverage (normal pillar errors do not make the geometric index exactly normal). Today's range: 28–67%. The band is close to symmetric around the headline today; it turns asymmetric only near the ends of the scale, where the logistic saturates.
Three honest caveats. First, this range replaces an earlier “±10pp” figure that was asserted but never derived. Second, the propagation assumes scoring errors are independent across pillars; with one analyst, shared OSINT sources, and wartime common shocks, that is optimistic — at a correlation of ρ = 0.3 the range widens to 23–73%. Read the published range as a floor with respect to what it excludes — cross-pillar correlation, parameter uncertainty (k, x₀, the weights) and model-structure uncertainty. Third, it is not a floor on the upside: 3 pillars currently sit at 99, where upside score error has almost nowhere to go, so under any error model that respects the 0–100 score bounds the distribution of recomputed headlines shifts downward — its center falls a few points below the published headline and its upper tail falls short of the published upper endpoint. The bounded scale pulls the top of the range in even as the excluded uncertainty sources push the bottom out.
About 80% of the range's variance comes from a single pillar: security apparatus cohesion. In plain terms, the band mostly measures how sure we are about IRGC cohesion.
This table is generated by the same code module that computes the headline: the exact effect (a recomputation, not an approximation) of each pillar's raw score moving +1 point, and how far the headline would move if that pillar alone went to its worst-for-regime extreme.
| Pillar | Weight | Score | +1 point → RC-1 | Worst-case extreme |
|---|---|---|---|---|
| Military Degradation | 15% | 99 | +0.6pp | +0.6pp |
| Security Apparatus Cohesion | 25% | 43 | −1.8pp | +44.1pp |
| Economic Fragility | 15% | 99 | +0.6pp | +0.6pp |
| Popular Mobilization | 12% | 96 | +0.5pp | +2.0pp |
| External Environment | 10% | 99 | +0.4pp | +0.4pp |
| Information & Narrative Control | 8% | 9 | −0.4pp | +3.1pp |
| Regime Adaptive Capacity | 15% | 21 | −0.8pp | +14.6pp |
Evaluated at today's values (CPI 82.5, headline 48%); valid for small moves. Inverted pillars carry a negative sign: a higher score means more regime strength and less pressure.
Saturation deserves plain words. Four direct pillars sit at 96–99 — near ceiling, with almost no room to add pressure. Until the Day-136 recalibration that pinned the headline high on a flat stretch of the curve; today the headline sits in the curve's steep zone — sensitivity is about 5pp of RC-1 per CPI point — so multi-point daily moves in either direction are expected behavior when cohesion evidence shifts, not a malfunction. The structural ceiling of the headline remains 96.8% at CPI = 100 — the model never declares certainty.
The model's parameters — the pillar weights, the logistic midpoints (83, 70) and slopes (0.2, 0.15) — were set by structured expert judgment and checked for consistency against ten historical episodes of authoritarian regimes under acute military and economic pressure. They were not statistically estimated: with roughly eleven free parameters and ten cases scored with outcomes already known, no fit statistic (classification accuracy, Brier score) would be meaningful, and we no longer publish one.
• Full regime change: Iran 1979, Iraq 2003, Libya 2011, Afghanistan 2001, Syria 2024, Romania 1989
• Ruler fell, state apparatus survived: Egypt 2011 — Mubarak exited, but SCAF preserved the military state; Any Transition yes, RC-1 no. That is exactly the distinction the site's two curves draw.
• Regime survived: Iraq 1991, Syria 2011–15, Venezuela 2019
Iraq 1991 is the case that disciplines this set the most: catastrophic military defeat, uprisings in 14 of 18 provinces, economic collapse — yet the Republican Guard held together and the regime survived twelve more years. It is why security cohesion carries the heaviest weight (25%), and why on Day 136 the index's aggregation became conjunctive (a weighted geometric mean); a cohesion-specific gate remains a specified option for the future. Two collapse pathways appear in the set: external conquest (Iraq 2003, Afghanistan 2001, Libya 2011) and internal defection cascade (Iran 1979, Romania 1989, Syria 2024). Absent a ground invasion, Iran 2026 is on the second path — which runs through the security apparatus.
An earlier version of this page listed Iraq 1991 under “collapsed” and claimed “10/11 classification accuracy”; both were wrong — Saddam’s regime survived until 2003, and no accuracy metric is estimable at this sample size. Corrected on Day 136; see the Integrity Log.
Day 93 recalibration (May 31, 2026). A note on integrity. Until this release, the headline RC-1 probability was set manually in the model's data file rather than computed from the pillar scores beneath it. The methodology page described a six-step pipeline — multipliers, a rally dampener, an “RC-1 fraction” — but that pipeline was never actually running. For 73 days, the number you saw was an editorial judgment dressed up as a computation. As of Day 93, the formula described on this page runs end-to-end on every update: CPI is computed from the seven pillar scores, and RC-1 and Any Transition are derived from CPI via the logistic curves. Legacy parameter fields have been retired, not patched. The switch to a live formula produced a small step-change in the headline when it went live; that reflects the gap between human judgment and the formula's output, not a sudden shift in conditions on the ground. We are publishing this caveat openly because the credibility of the model depends on it.
Day 136 update (July 13, 2026). The Day 93 fix relied on discipline: the formula ran on every update, but a person still copied its output into the site's data file. As of today that manual step no longer exists. The site derives every published figure — CPI, RC-1, Any Transition, the share image and its cache key — directly from the seven pillar scores in code at build time. The stored headline field has been deleted from the data files, and an automated pre-publish check rejects any release where the two language editions disagree or the chart's latest point differs from the computed headline. Every number published since Day 93 was formula-driven; what changed today is that this is guaranteed by the code rather than by procedure. Also corrected today: the historical-cases section listed Iraq 1991 as a collapse (Saddam's regime survived until 2003) and cited a meaningless accuracy statistic, and the Persian edition's chart was missing 24 days of history (May 15 – June 7), now restored. What Day 93 made a promise, Day 136 makes a property of the software.
Day 136 — an open scoring question. Publishing a scoring rubric cuts both ways, and we will not pretend otherwise. Under the rubric now published for security apparatus cohesion, the recent evidence — four consecutive nights of coordinated multi-front IRGC retaliation, and enforcement of a declared Hormuz closure — reads closer to the 25 anchor (“retains coordinated strike capability but can no longer secure regime rituals”) than to the score of 11 held at the time, which priced in the March structural wounds: a decapitated command layer, coercion-dependent allegiance to an invisible leader, and confirmed desertion waves. The stakes are material, and we would rather state them than have a reader compute them: under the then-current formula, re-scoring cohesion to 25 would have put the headline near 73%; to 45, near 50%; to 55, near 38%. The score is under review. If it changes, the change will be published as a dated recalibration with the reasoning — not slipped into a daily update.
Day 136 recalibration — v2.0 (July 13, 2026). The open scoring question above is resolved the way this page said it would be. We convened six independent raters with deliberately different methods — two rubric literalists, a civil-military theorist, a military-operations analyst, an OSINT source-skeptic, and an Iran base-rates specialist — each blinded to the previous score, to the headline, and to the consequences, and asked each to score security-apparatus cohesion against the published rubric from the published evidence. Their scores: 38, 38, 40, 42, 43, 44 — median 41, with every rater's 80% interval overlapping in 33–48. Their shared finding: the evidence meets the 50-anchor's clauses almost fully (loyalty held by threat, force regeneration failing, command still executing major operations) while the 25-anchor's defining markers — whole-unit defections, loss of territorial control — have not occurred. The score moves from 11 to 41, the panel median. Two further changes ship with it: the CPI is now a weighted geometric mean rather than a weighted sum — the conjunctive form the historical record demands, with no new parameters in the headline computation (the weights and both logistic curves are unchanged; the per-pillar observability sigmas that drive the uncertainty band are new, and listed in the Uncertainty section) — and the headline now carries an explicit 12-month window with derived 90-day and 6-month companions, making it scoreable for the first time. Net effect: the headline moves from 85% to 47% — roughly 30 points from the cohesion rescore and 8 from the aggregation change (an order-dependent decomposition). Nothing happened on the ground today; what changed is the measurement of the least observable quantity in the model, the one that dominates the uncertainty band the site now publishes. The correction also lands far outside the 74–91% band that scoring-noise propagation would have drawn around the old headline — the site previously published only an asserted ±10pp figure. Bands of that kind model noise, not a misclassification of this size; that is exactly why the Uncertainty section labels its band a floor. Honest limits of the audit: the raters were AI raters with distinct analytical personas rather than independent human experts, and the evidence packet was assembled by the same process being audited. The previous headline is shown alongside the new one for two weeks, and the chart marks today's discontinuity. This is what the rubric was published for.
Day 143 review release (July 20, 2026). An adversarial review of every published statistic — run against this site's own claims — found four things worth putting on the record. First, the Day 93 entry above misdates its own event: git history shows the formula release actually shipped on June 8, in a batch that also rewrote the May 31 reading from the 89% it had displayed to the formula's 81% and computed the June 1–7 points retroactively. The frozen entries above stay as written; this entry corrects the record, and the history table now labels the pre-v2 era accordingly. Second, the chart note claimed the v2 formula “barely moves the history” — false over the hand-set era, where the recomputed line diverges from the published one by ~8 points on average and up to 54 early in the war; the note now states the divergence. Third, uncertainty language has been tightened: coverage is approximately (not “exactly”) 80%, and the band's “floor” framing is now scoped — near the score caps the published upper endpoint overstates rather than understates, because bounded scores can only err downward from 99. Fourth, Any Transition has been brought to parity with the headline: it now carries the same 12-month window, its own 80% range, explicit resolution baselines (including that the March 2026 succession does not resolve it), and a published grading protocol naming the Brier rule. One commitment: the Day-136 blind panel audited only the cohesion score; the same six-persona protocol will now be run on the other six pillars, and the result will be published as a dated recalibration — whatever it finds, in either direction.